Experimental Demonstration of Conversion-Based SNNs with 1T1R Mott Neurons for Neuromorphic Inference

SNNs using the conversion-based approach could benefit the energy efficiency of inference and retain high accuracy of DLNs. However, transistor-based spiking neurons and synapses are not scalable and inefficient. In this work, a Mott neuron with 1T1R structure is designed to meet the requirement of the conversion-based approach, whose spiking rates dependence on voltage naturally implements the rectified linear unit (ReLU). Based on the 1T1R Mott neuron, we experimentally demonstrated a one-layer SNN (320 ×10), which consists of RRAM synaptic weight elements and Mott-type output neurons, for the first time. Attributes to the rectified linear voltage-rates relationship of the 1T1R neuron and its inherent stochasticity, 95.7% converting accuracy of the neurons and 85.7% recognition accuracy in MNIST datasets are obtained. At last, a neuron X-bar architecture is proposed for parallel multi-tasking and better system integration.